Skip to content

Add Emotions to a Voice with Delivery Modes

Make a character sound angry, afraid or joyful, and switch between moods while the game runs. A voice profile holds one voice and a list of delivery modes: named ways of speaking, such as angry or whisper. Game code switches the mode with SetAgentDeliveryMode, and the next sentence is spoken in the new mood.

With a cloning model (Qwen3-TTS Base or Pocket TTS), each mood is a reference clip of the same character speaking in that mood. This page shows how to make those clips, put them in a profile, and switch between them.

Prerequisites

Why one clip per mood

Cloning copies the register and energy of the reference clip along with the voice. In our measurements, neutral text cloned from a shouted, angry clip came out raised and loud on both models. Cloned from a sad clip, it came out low and flat. Cloning models have no emotion parameter: setting a delivery mode's Instruction on them fails with 3023. So the clip is the mood:

  1. One clip per character per mood, for example voices/guard/neutral.wav, voices/guard/angry.wav, voices/guard/afraid.wav.
  2. The neutral clip in the profile's Reference Voice. That is the built-in neutral mode.
  3. One delivery mode per mood in the same profile, each with its own Reference Voice.
  4. SetAgentDeliveryMode("angry") at runtime. It needs no agent rebuild.

Every mood must sound like the same person. A clip from another speaker changes who is talking, not just how.

1. Plan the moods

  • Pick the few moods your game uses. Each one is a clip to make and to check. Three to five per character is plenty for most NPCs.
  • Use the same mode names on every character. Then one call, such as SetAgentDeliveryMode("angry"), works on everyone. A mode with no fields set sounds like neutral, so a character without an angry clip can still define an empty angry mode.
  • Keep names short and lowercase, such as angry, afraid, joyful. Game code and the editor's dropdown use them as written.

2. Make one clip per mood

There are two ways to get a set of clips that sound like one person.

Record one actor

This gives the best result. Record every mood with the same actor, in the same session, with the same microphone and room. Follow the clip rules in Clone a Voice → Record a reference clip. Ask for each mood in character, and keep the same distance to the microphone, even when the actor shouts or whispers.

Design the moods with Qwen3-TTS

Without an actor, render each mood with Qwen3-TTS VoiceDesign. The steps for one clip are in Design a Voice Clip with Qwen3-TTS. For a set of moods:

  1. Split the description into identity and delivery. The identity (gender, age, pitch, timbre, accent) stays word for word in every mood. Only the delivery changes.
  2. Write a line in each mood. An angry line for the angry clip, a frightened line for the afraid one.
  3. Render several takes per mood and keep the ones that sound most like the neutral clip. Listen to them side by side. Drop a take that sounds like someone else, even if its mood is right.

For example, a palace guard:

Mood Voice description
neutral A man in his forties with a deep, slightly hoarse voice and a light Northern English accent. He speaks calmly, at an even pace.
angry A man in his forties with a deep, slightly hoarse voice and a light Northern English accent. He is furious and shouts, fast and sharp.
afraid A man in his forties with a deep, slightly hoarse voice and a light Northern English accent. He is frightened; his voice shakes and he speaks quickly and quietly.

Each render is a new sample, so designed moods can drift slightly away from each other. Picking takes by ear keeps the drift small.

Put the clips together in one folder per character:

<storage root>/
  └─ voices/
      └─ guard/
          ├─ neutral.wav
          ├─ angry.wav
          └─ afraid.wav

3. Add the modes to the voice profile

  1. Select the character's Voice Profile asset (Create → Tryll → Voice Profile if it has none).
  2. Set Tts Model Name to the cloning model and Reference Voice to the neutral clip.
  3. Under Delivery Modes, add one element per mood. Set its Name (angry) and its Reference Voice (voices/guard/angry.wav, with the … button or typed). Leave Instruction empty.
  1. Open the character's Tryll Voice Profile data asset.
  2. Set the TTS model to the cloning model and Reference Voice to the neutral clip.
  3. Under Delivery Modes, add one element per mood with its Name and Reference Voice. Leave Instruction empty.
from tryll_client import DeliveryMode, VoiceProfile

guard = VoiceProfile(
    tts_model_name="Qwen3-TTS 12Hz 1.7B Base (Q4_K_M)",
    reference_voice="voices/guard/neutral.wav",
    delivery_modes=[
        DeliveryMode(name="angry",  reference_voice="voices/guard/angry.wav"),
        DeliveryMode(name="afraid", reference_voice="voices/guard/afraid.wav"),
    ])
agent = client.create_agent(graph, voice=guard)

Tryll::VoiceProfileT guard;
guard.tts_model_name  = "Qwen3-TTS 12Hz 1.7B Base (Q4_K_M)";
guard.reference_voice = "voices/guard/neutral.wav";

auto angry = std::make_unique<Tryll::DeliveryModeT>();
angry->name            = "angry";
angry->reference_voice = "voices/guard/angry.wav";
guard.delivery_modes.push_back(std::move(angry));

auto afraid = std::make_unique<Tryll::DeliveryModeT>();
afraid->name            = "afraid";
afraid->reference_voice = "voices/guard/afraid.wav";
guard.delivery_modes.push_back(std::move(afraid));

Tryll::Client::AgentCreateOptions opts;
opts.voice = std::move(guard);
auto agent = client.CreateAgent(graph, std::move(opts));

A mode can also set Speed Scale, for example 1.15 to make afraid a little faster.

By default, the server prepares every clip in the profile when the agent is created, so the first sentence in a new mood has no extra delay. Set Prepare to Lazy to prepare each clip on its first use instead. A bad or missing clip path, in the profile or in any mode, fails agent creation with 3023, and the message names the field.

4. Switch the mood at runtime

await agentComp.SetDeliveryModeAsync("angry");    // the guard loses his temper
await agentComp.SetDeliveryModeAsync("neutral");  // and calms down again

Call Set Delivery Mode on the Tryll Agent component with the mode name. It reports completion through its delegate.

agent.set_delivery_mode("angry")
agent.set_delivery_mode("neutral")

agent.SetDeliveryMode("angry", 1.0f);
  • It applies from the next sentence. The call is accepted at any time, even mid-turn. A sentence already sent to the TTS model keeps its mood.
  • neutral is built in. It always exists and means the profile's own Reference Voice.
  • Intensity does not blend clips. A clip is all or nothing. On Pocket TTS, intensity only scales the mode's Speed Scale; on Qwen3-TTS Base it does nothing.
  • An unknown name fails with 3024 UnknownDeliveryMode, and the current mode stays.

To try a mood without code, use the agent component's Voice Delivery override in the Inspector (Unity) or Details panel (Unreal). Its dropdown lists neutral and the profile's modes, and changing it on a live agent switches the mood at once.

Decide when to switch

Tryll does not detect the mood for you: only the API switches the mode, and the language model cannot. Your game owns the character's emotional state, the same way it owns health or quest progress, and calls SetAgentDeliveryMode when it changes. Pass the same mood into the prompt too, so that what the character says matches how it sounds. See Talking NPC → Emotions.

Example: the Unity demo

The Unity demo's EmotionalVoiceCharacter is a complete example. Kira, a K-pop singer giving a magazine interview, speaks with one Qwen3-TTS voice and eight moods:

  • Voice profile: Assets/Voices/VoiceCharacterWithEmotions, on Qwen3-TTS 12Hz 1.7B Base (Q4_K_M).
  • Neutral clip: voice/young_female/female_young_calm.wav.
  • Delivery modes: angry, very_angry, fear, hysteria, joy, sad, flirty and terror, each with its own clip in TryllStorage/voice/young_female/.
  • Workflow: Assets/Workflows/EmotionalVoiceCharacter.

The clips were made with Qwen3-TTS VoiceDesign, with one description per mood. To hear it, open DemoScene, select the EmotionalVoiceCharacter object, pick a mood in the agent component's Voice Delivery dropdown, then enter Play mode and talk to Kira. Change the mood while she speaks to hear the switch at the next sentence.

Other TTS models

How a mode expresses a mood depends on the model:

Model What a mood is
Qwen3-TTS Base, Pocket TTS A reference clip per mood (this page)
Maya1 A text Instruction, such as angry or whisper, with intensity
Supertonic 3, Orpheus No emotion control: a mode can only change speed, or stay empty

The details for each model are in Voice Profile → Delivery modes.